I build a lot of computer vision systems and did my MSc on classifying what’s happening in video and I am VERY skeptical that it’s possible to prevent this from happening (without having a person watch every video).

You can edit a video to make it appear as safe content and it’s a cat and mouse game because you need humans to annotate those variations before a ML system can detect them (an unsupervised model will likely lead to loads of false positives).

I think the solution is better control over what we see and who we trust enough to allow them to broadcast into our brains.